Auto-decode columns declared as JSON in output modes (issue #579)

Columns declared with type JSON now decode automatically wherever nested
JSON is already supported (query/rows/search/memory output, and the
Python API), without needing --json-cols. Uses sqlite3's built-in
detect_types=PARSE_DECLTYPES + a registered JSON converter, so it works
transparently through joins and aliases. Flat formats (CSV/TSV/table/raw)
re-serialize decoded values back to JSON text instead of a Python repr.
Also adds "JSON" as a first-class column type for create()/add_column()/
transform(), and fixes column_affinity("JSON") so transform() doesn't
rewrite JSON columns as REAL.

tests/test_column_affinity.py

@@ -32,6 +32,9 @@ EXAMPLES = [
     ("BOOLEAN", float),
     ("DATE", float),
     ("DATETIME", float),
+    # Not a real SQLite affinity, but sqlite-utils treats it as TEXT so
+    # transform() does not mangle JSON columns - see issue 579
+    ("JSON", str),
 ]